Integrations

Clusy works with a few outside services on your behalf: somewhere to version your notebooks, somewhere to publish them, warehouses the agent can query, and hubs it can pull data from. You manage them all under Settings → Integrations, where you connect, disconnect, and check the status of each.

Each one is a door. Your credential is what opens it, and the agent then works through it on your behalf — in both directions, sending work out and pulling data back in.

You only connect what you need. None of this is required for the everyday work of building and running notebooks, and none of it is required for the agent to search the web or find a public dataset. An integration you haven't connected isn't broken; it's just shut, and you can open it whenever you like.

An integration is a door: your credential opens it, and the agent then works through it in both directions — data in as well as out. Leave one unconnected and it is simply shut, not broken; web research needs no door at all.

GitHub

Connect GitHub to version your notebooks as readable diffs. You choose which repositories Clusy can reach rather than handing over your whole account, and once a project is linked to a repo the Git tab in the sidebar comes alive. See GitHub for the full workflow.

The same Settings section holds your Git configuration, including whether pushes should always force-push. Leave that off unless you know why you want it.

Hugging Face

Hugging Face connects over OAuth, so you authorize it once. It works in both directions.

Reading. The agent can search the Hugging Face Hub for datasets and models, read their cards, and pull one into the workspace. It does this for public repos with nothing connected at all; a connection is what gets it into your private repos.

Publishing. With Hugging Face connected, you can publish a notebook along with the models or datasets that go with it to a repo under your account.

If OAuth publishing isn't available, you can instead paste a Hugging Face access token, which lets the agent search and read Hub repos (including your private ones) even though publishing stays off.

Kaggle

The agent can already search Kaggle's public datasets and read their metadata without any connection at all. What a connection buys you is publishing: a Kaggle API key (the username and key from your kaggle.json) lets you push a notebook as a Kaggle kernel and upload the models or datasets that go with it. See Publishing.

Databricks

Connect Databricks and the agent can query your lakehouse directly: list catalogs, schemas, and tables, describe a table, and run SQL against a SQL warehouse, bringing the results into the notebook. You connect with your workspace host and a personal access token from your Databricks account.

Snowflake

Snowflake works the same way for your Snowflake warehouse: the agent can list databases, schemas, and tables and run SQL, using your account identifier and a programmatic access token you create in Snowflake.

With either warehouse connected, you can ask things like "pull last quarter's orders from the warehouse and load them into a DataFrame" and the agent handles the query itself.

Firecrawl (optional)

You do not need to connect anything for the agent to use the web. Searching, fetching a page, and extracting structured data from one are built in and work out of the box, as does research against arXiv and OpenAlex.

Firecrawl is there for people who already pay for it and would rather the crawling ran on their own quota, with their own rate limits and their own account. Paste your key and the agent's Firecrawl-backed scraping, search, mapping, and file parsing run against it.

If you skip this, nothing turns off.

Where data comes from

Two of these integrations are about getting data in, and it's easy to miss:

• Ask for a public dataset and the agent searches the Hugging Face Hub and Kaggle unprompted, ranks what it finds, and loads the one you pick. See Find a dataset.
• Ask a research question and it searches arXiv and OpenAlex, screens the papers, and writes up what it found with citations. See Research.

Neither needs a key. What connecting Hugging Face adds is your private repos; what connecting Kaggle adds is the ability to publish back.